(Original title: artificial intelligence assisted screening for lung cancer: This company wants to help doctors to lighten their burdens and ease the conflicts between doctors and patients)
News reporter Wang Xinxin
“In China, 70% of the diagnosis and treatment need to be diagnosed based on the information given by the medical images. And this system is still constantly developing. The volume of images and images is about 40% annual growth, but correspondingly, the number of image doctors per year At most, it is only increasing at a rate of 3%, so China is now very short of imaging doctors.†Liu Shiyuan, director of the Radiology Department of Changzheng Hospital of Shanghai Second Military Medical University and deputy director of the Chinese Radiological Society, told Xinhua News ().
The lack of imaging doctors raises another problem that leads to doctors' overwork. When doctors work in overload, the probability of making wrong judgments will increase, and even some early-stage diseases that can be found may be missed. Such a vicious cycle has become a driving force for the tension between doctors and patients.
At present, the key to solving this problem lies in the artificial intelligence technology. In 2013, the MIT Technology Review rated one of the top ten technological breakthroughs that automatically identified diseases and improved deep learning in hospital diagnosis. At present, the application of artificial intelligence in the fields of finance, medical care and manufacturing is rapidly growing. McKinsey estimates that by 2025, the total market for AI applications will reach $127 billion. Another agency, IDC, predicts that the annual compound growth rate of the domestic medical informatization solution market from 2012 to 2016 will reach 14.3%. In the future, this market is expected to exceed 30 billion yuan.
In China, training models using deep learning techniques can help doctors identify lung lesions and cancer in clinical practice. Changzheng Hospital of Shanghai Second Military Medical University became one of the earliest hospitals to embrace this technology, and the company that they worked with was a start-up company that had no interest before 2016—think science and technology.
“When I returned to China in 2014, I discovered that in the domestic medical field, many doctors are actually overloaded every day. When I was determined to do this, I discovered that actually there were many people around the hospital because of the hospital’s Missed or misdiagnosed missed the best treatment time.†Chen Kuan said in an interview with Yu Xin News ().
Using Propagation Data to Teach Machines to Identify Lesions
However, it is easy to talk about and it is difficult to do.
“Our cooperation with Chen Kuan began in October 2016. To be honest, the results were not optimistic at the beginning of cooperation. Although the machine can detect lung lesions, many of them are false positives,†said Liu Shiyuan.
The reason why the application of medical image recognition and diagnosis in the lungs first has a certain scientific basis. Because the lung itself is an organ containing air, its medical images have the best natural contrast. If there is a lesion in the patient's lungs, it is equivalent to having a white point on the black paper, which is obvious at a glance. For the machine, it is easier to identify. The medical images of other organs, such as the liver and brain, have a white point if there is a lesion on a gray paper, and the contrast is not strong.
Liu Shiyuan made a metaphor for the false positives that were detected by the machine. The false positive detected by the machine is equivalent to a police officer identifying ten bad people in the crowd, but found that eight are good people. Another problem is the possibility of missed diagnosis. In other words, the police can identify murderers and other felony criminals, but can not identify the petty criminals of small thieves. For lung lesions, many lung nodules are initially small and light and take time to grow. However, for many patients, this is the best time to visit.
To change the above situation, high-precision medical data integrates the key. Yang Lekun, director of the Facebook Artificial Intelligence Laboratory, once explained the image recognition technology of machine learning. For the machine to identify what is in the picture, first the human needs to tell it what is in the picture through a large number of marked pictures. Liu Shiyuan collaborated with Science Technology and used the data set clearly marked by the hospital to train the model to improve the accuracy of machine identification.
“This is equivalent to asking a teacher to teach an ignorant child to read or write. If you ask a teacher who is not reliable and give him the wrong information, the child will learn more and more badly; on the contrary, if you ask a reliable teacher, He can learn to be smarter," said Liu Zhuren.
With the help of a complete set of lung standardization reconstruction and evaluation system at Changzheng Hospital, after about 6 months of training and experimentation, the currently recognized technology products have been improved in the recognition rate of lung lesions. Of the 10 "criminals" identified, the machine's error rate has dropped to one or a half. At the same time, compared with young doctors in the hospital, the machine found that the sensitivity of nodules was about 50% higher than that of human eyes. A nodule that humans may miss in the naked eye, the current machine has not missed.
The hardship behind the high recognition rate
It is encouraging to think that the achievements made by science and technology in the Long March Hospital are encouraging, but there are still a few hospitals that are willing to embrace new technologies in China. Most of the hospitals are still waiting to see the stage.
“Using artificial intelligence to solve some problems in the medical field is still a relatively early stage, or it has just begun. We have already felt that artificial intelligence is the fourth industrial revolution and it is a huge opportunity. There are many companies In this work, for hospitals or doctors, only some people see the prospects and feel that they are worthy of attention. Unfortunately, doctors who work and explore are still a minority, and most of them are still watching.†Liu Shiyuan said.
Chen Kuan encountered the same problems in the early days of returning to China. He told the Xinhua News (): “Actually, many challenges encountered on the road to entrepreneurship are normal. The biggest pressure that has given me the most pressure is not a technical issue. More It was when I went to the hospital to talk about cooperation. The doctors I met didn't accept it. It made me have nothing to do."
In addition to the traditional concept of new technology is not yet recognized, another blocker comes from the data. For deciding technology, in order to make its model more accurate, it needs cooperation with domestic first-rate hospitals to get reliable data. On the other hand, how to ensure data security also needs consideration. In cooperation with hospitals, we believe that technology uses AI-DR, AI-CT and AI-Scholar, and Beijing Union Medical College Hospital. Shanghai Changzheng Hospital and Wuhan Tongji Hospital cooperated.
According to Liu Shiyuan, the multi-center lung cancer screening research currently under the leadership of himself is conducting statistics and screening of lung cancer data for Shanghai residents over the age of 40. More than 16,000 low-dose pulmonary nodules in the study were desensitized and “fed†to the machine. Desensitization refers to the removal of patient's personal information, leaving only what the machine needs. With such research and accurate data, the machine has been greatly improved in the accuracy of lung cancer identification at a later stage.
For the machine, it is quite fortunate to have the scientific research data for training. More often than not, the data from clinical consultations with hospitals is more and more complex, but it is not very useful. If this is the case, doctors will need to accurately mark the machine before entering it. This is a very arduous process for hospital doctors. Because each patient's image is stored as normal, there are more than 200, and even some may have four or five hundred images.
"These images need someone to take a serious look and filter. After reading it, there are lesions, abnormalities, and markings. It is also a very laborious process to ensure that the marks are correct. Our experienced head doctor looks at a patient. It may take at least half an hour or so, and you need to think about how long we would have to spend on marking 6,000 patients.Some doctors need to sit in front of their computer for a long time and see how giddy they feel. "Liu Shiyuan said. He also used the child's education as a metaphor to explain the hard work of data marking. “This is like a young mother with a child. The last child was admitted to Yale and Cambridge. The outsider saw the end result but didn’t know the process was hard. When the product was on the line in the hospital, the patient’s condition could be seen in about one second. , but the actual early need to do a lot of meticulous work."
AI allows doctors to watch movies and balance medical resources
As the saying goes, sharpening does not mistakenly cut the firewood. Although the labeling and input of early-stage data requires a lot of time and effort, the trained model can bring a tremendous improvement to the hospital's diagnosis and treatment. The first is to liberate the labor force of many doctors. Using machines instead of doctors to identify medical images can reduce the time doctors spend watching movies, freeing up more medical resources. Let doctors have more energy to communicate with patients.
“A person is not a machine, he will be tired, but also go to the toilet and eat. When he is too tired, the efficiency of doing things will fall, and then there may be missed or even misdiagnosis. If there is misdiagnosis, missed diagnosis, there may be medical In addition, in fact, the medical imaging department's job is to look at images, write reports, and describe illnesses. This is a very boring and repetitive process," said Liu Shiyuan.
Second, the combination of artificial intelligence and medical care can also solve the medical resources in backward areas in the future. It is assumed that the cooperation between the technology and the Long March Hospital continues, but the two systems that jointly create a diagnosis and treatment system will ultimately form products and experiences, which will be used by hospitals in remote areas.
“Now the state advocates grading diagnosis and treatment and promotes the sinking of quality medical resources. The quality medical resources of Kitasato are not likely to go to Qinghai and Tibet, because it is unrealistic for doctors to abandon their lives in North Shanghai to work in these places. But if it’s because of our experience, And the cooperation with the deliberation has formed a good artificial intelligence product. Using this product in the Alibaba District People's Hospital in Tibet may have the same result, which is equivalent to the fact that Ali’s side can also achieve the diagnosis of the Long March Yard. The level is a huge gospel for the people of the country," said Liu Shiyuan.
Third, artificial intelligence can also improve hospital management efficiency. According to Chen Kuan, apart from being able to use medical imaging in artificial intelligence, it is also possible to integrate all aspects of disease in the future. In addition, from the point of view of hospital management, digging patient data, financial management, bed occupancy rates, turnover rates, etc. is very effective for the management of hospitals. According to Xinhua News, Changzheng Hospital introduced Cisco China's precision medical platform to help support the basic calculation of artificial intelligence, improve the hospital's intelligence and data management, and enable patients to enjoy better services.
Finally, with the increasing role of artificial intelligence in the medical field, it has also led to discussions: Will it replace the doctor in the future?
"The machine is superior to human beings in the study of massive data, memory and superior computing power. But the human brain is connected by complex chemical transmitters. Now human understanding of the working principle of the brain is just an iceberg. The machine is only part of the mechanism that imitates the human brain. From this point of view, it is impossible or even impossible for artificial intelligence to replace human beings. Finally, there is a point that doctors It is a profession. It needs to get a doctor's license, there are some ethical restrictions, there is a threshold, these are machines can not do." Liu Shiyuan told the news ().